Improve Aerospike Database performance and predictability by leveraging Intel® Ethernet 800 Series Network Adapters with Application Device Queues (ADQ)

Improve Aerospike Database performance and
predictability by leveraging Intel®
Ethernet 800
Series Network Adapters with Application
Device Queues (ADQ)
In Dell EMC PowerEdge R740xd servers, enabling ADQ boosted
Aerospike performance and led to near-linear scaling in both
DRAM and Intel®
Optane™
persistent memory configurations
Scale-out applications that rely on communication between
servers depend on fast networking performance. Now, a feature
of Intel®
Ethernet 800 Series Network Adapters, Application
Device Queues (ADQ), offers a streamlined path for data access.
In the Principled Technologies data center, we set up three
Dell EMC™
PowerEdge™
R740xd server nodes and used a
built-in benchmark tool to assess Aerospike Database 5 (a
NoSQL database platform) performance with ADQ enabled
and ADQ disabled at multiple node counts. Our tests showed
that enabling ADQ with the Intel Ethernet 800 Series Network
Adapter provided a significant improvement in throughput,
latency, and predictability for Aerospike databases—which helps
deliver responsive experiences for end users. Enabling ADQ also
offered near-linear scaling for Aerospike Database applications as
we added server nodes.
To make the most of your Aerospike Database workloads on Dell
EMC PowerEdge R740xd servers, enabling Application Device
Queues on Intel Ethernet 800 Series Network Adapters can give
these high-priority database applications a boost.
Near-linear scaling
for Aerospike databases
Increase Aerospike
Database throughput
by up to 64%
Increase predictability
of response times by
up to 41%
on Aerospike databases
April 2021
Improve Aerospike Database performance and predictability by leveraging Intel
Ethernet 800 Series Network Adapters with Application Device Queues (ADQ)
A Principled Technologies report: Hands-on testing. Real-world results.

The importance of being predictably fast
Distributed databases—where data resides on more than one node, typically many physical nodes—rely on more
than just server processing power to ensure fast performance; the network connecting this data must be fast.
Database performance must be fast, but also have predictable response times, limiting slow outlier requests that
add up to latency problems as the applications scale. Dell EMC PowerEdge R740xd servers utilize Intel Ethernet
800 Series Network Adapters to meet these twin challenges.
Intel Ethernet 800 Series Network Adapters offer Application Device Queues, which route data to dedicated
hardware queues to decrease latency, improve response time predictability, and ultimately deliver higher
throughput for your critical applications. By creating what are essentially express lanes for high-priority data,
enabling ADQ on the Intel Ethernet 800 Series can help you more easily meet service-level agreements (SLAs)
and scale applications confidently while providing very low latencies.
To show how ADQ can benefit applications, we created a testbed with three Dell EMC PowerEdge R740xd
servers and compared Aerospike Database performance with ADQ enabled and with ADQ turned off (see
Figure 1). (To see other applications that can utilize ADQ, visit www.intel.com/adq.) We tested the Aerospike
Database application using a benchmark built into the Aerospike platform: the Aerospike C benchmark. This
simple benchmark allows users to insert records into the database and run tests with a variety of read/write mixes
(we used 100% reads in our tests) and thread counts to assess overall Aerospike Database performance. The
benchmark reports its results in transactions per second (TPS). For complete testing details and results, read the
science behind the report.
About Aerospike Database 5
In our tests, we used Aerospike Database Enterprise Edition build 5.3.0.2. Aerospike Database 5 is a NoSQL database
platform that supports multiple deployment options on premises or in the cloud. According to Aerospike, “[the database
platform] supports two distinct active-active deployment models, multi-site clustering (MSC) and cross datacenter replication
(XDR) for building a global data hub.”1 To learn more about Aerospike Database, visit https://www.aerospike.com/products/
database-platform/.
Figure 1: Our testbed setup for Aerospike Database testing. Source: Principled Technologies.
April 2021 | 2

Figure 2: Aerospike database read throughput performance, in transactions per second, of the Dell EMC PowerEdge
R740xd at multiple server node counts with Intel Ethernet 800 Series Network Adapters with ADQ on and ADQ off.
Source: Principled Technologies.
Throughput and scaling
Throughput is a measure of how much work can be done in a certain time—how many orders, operations,
Megabytes, and so on. The more application throughput a server can handle with good response times, the
more customer requests it can process before you need to support additional nodes. In tests of Aerospike
databases, we compared the application throughput with ADQ enabled vs. ADQ disabled to show how using
ADQ can prevent network bottlenecks and enable more customer requests to get through. Both ADQ enabled
and disabled configurations use optimization via NUMA pinning.
As Figure 2 shows, enabling ADQ dramatically improved the number of random read transactions per second
that the Dell EMC PowerEdge R740xd could handle at multiple server node counts, increasing read throughput
by up to 64.8 percent.
0
10,000,000
20,000,000
30,000,000
40,000,000
50,000,000
1 server 2 servers 3 servers
Aerospike database read throughput performance
with and without ADQ
29,169,294
43,217,907
17,971,014
26,219,536
15,051,776
10,254,738
Transactions
per
second
(Higher
is
better)
ADQ-on ADQ-off
About 2nd
Generation Intel Xeon Scalable processors
The 2nd
Generation Intel Xeon®
Scalable processor platform features a wide range of processor types to support the
workloads you run, including Bronze, Silver, Gold, and Platinum. According to Intel, the 2nd
Generation Intel Xeon Scalable
platform can handle a variety of workloads, including enterprise, cloud, HPC, storage, and communications. The new
processor line also supports a new memory and storage technology to further accelerate workloads, Intel Optane DC
persistent memory. To learn more about the Intel Xeon Scalable processor family, visit https://www.intel.com/content/www/
us/en/products/docs/processors/xeon/2nd-gen-xeon-scalable-processors-brief.html.
April 2021 | 3

0.0
0.5
1.0
1.5
2.0
2.5
3.0
Aerospike database read throughput performance scaling
1.75
1.00
Relative
transactions
per
second
(Higher
is
better)
2.87
2.56
ADQ-on ADQ-off
Perfect scaling
3.00
2.00
1.94
Figure 3: Scaling for Aerospike databases over multiple Dell EMC PowerEdge R740xd server nodes with Intel Ethernet
800 Series Network Adapters with ADQ on and ADQ off. We included a line representing perfect scaling to show how
close each configuration came. Source: Principled Technologies.
Ideally, an application would scale linearly as identical servers are added—meaning that two servers do exactly
twice the work of a single server, three servers do exactly three times the work of a single server, and so on.
Because servers communicating with one another requires some amount of overhead to account for their ability
to work in concert, perfect linear scaling is generally not possible. Choosing hardware and settings that closely
approach this ideal ensure that you make the most of your available resources. As Figure 3 shows, enabling
ADQ did more than just provide better throughput overall—enabling ADQ also improved scaling noticeably as
we added identical server nodes compared to the same servers with ADQ disabled. At three nodes, the ADQ-
enabled configuration deviated just 4.3 percent from linear scaling.
Using ADQ in conjunction with the Intel Ethernet 800 Series Network Adapters to route traffic to dedicated
hardware queues allowed the Dell EMC PowerEdge R740xd servers to spend more resources on ensuring high
throughput for the Aerospike Database application, which could, at large scale, ultimately lead to the purchase
and maintenance of fewer servers to support user requests.
Latency and predictability
Even the best performing servers with high throughput and low average latency experience outlier requests
that fall outside of acceptable response times. At low server counts, these slowest response times, known as
tail latency, are not likely to significantly affect user experience. However, in deployments at scale—with many
servers supporting the same application—these slow responses can reduce the ability of your organization to
meet SLAs and adequately support users. Reducing jitter, the term for the variation in latency, can help you
predictably scale Aerospike Database applications to ensure consistent performance. In our tests, enabling ADQ
in addition to NUMA pinning, which assigns memory and CPU usage of Aerospike databases to a single NUMA
node, helped to reduce jitter and ensure predictable, consistent performance.
Just 4.3%
deviation from
linear scaling
at three nodes
April 2021 | 4

Figure 4: Latency for 99% of requests at multiple Dell EMC PowerEdge R740xd server node counts with Intel Ethernet
800 Series Network Adapters with ADQ on and ADQ off. Source: Principled Technologies.
Figure 4 compares the response times for 99th percentile (P99) latency—which means the response time that
99 percent of all requests achieved. Lower response times can help enable better overall performance, and
having a lower P99 latency means that nearly all response times from all requests were lower, which improves
predictability at scale. With a single server node, turning on ADQ resulted in a 41.7 percent improvement in
latency predictability.
0
100
200
300
400
500
600
700
800
Predictability of read latency (99th percentile)
400
600
280
480
Microseconds
(Lower
is
better)
ADQ-on ADQ-off
440
680
Figure 5 on the following page presents a latency chart with histogram buckets comparing a single Dell EMC
PowerEdge R740xd server node with ADQ enabled vs. the same server with ADQ off. The y-axis shows the
percentage of requests, while the x-axis shows different latency thresholds in microseconds. For both solutions,
100% of requests were >40 microseconds. With ADQ enabled, there was a very low percentage of requests that
took longer than 280 microseconds (as the line approaches 0), while with ADQ off over 20% of requests were
longer than 280 microseconds. (Note: To see complete histogram charts for two- and three-node testing, visit
the science behind the report.) By weighting the number of responses in each histogram bucket, we determined
that enabling ADQ provided a 32.0 percent lower weighted average latency than a single server node with ADQ off.
Figures 4 and 5 show that with ADQ enabled, the Aerospike Database application experienced less variability
and delivered most of its requests at a much lower latency. By ensuring fast response times as well as
predictability, organizations can scale Aerospike databases with more confidence.
April 2021 | 5

0
20%
40%
60%
80%
100%
Percentage of read requests at various latency thresholds
>40us >80us >120us >160us >200us >240us >280us >320us >360us >400us >440us >480us >520us >560us >600us
ADQ-on ADQ-off
Latency thresholds in microseconds
Percentage
of
requests
(Lower
is
better)
Figure 5: Latency histogram buckets for a single Dell EMC PowerEdge R740xd server node with Intel Ethernet 800
Series Network Adapters with ADQ on and ADQ off. Source: Principled Technologies.
About Intel Ethernet 800 Series Network Adapters
According to Intel, the new Intel Ethernet 800 Series offers technologies
that “address a variety of workloads used in Cloud, Communications, and
Enterprise market segments.”2
In addition to ADQ, 800 Series Network
Adapters support two RDMA storage transport protocols, iWARP and
RoCEv2, to offer storage flexibility for various business needs. To learn
more about the Intel Ethernet 800 Series, visit https://www.intel.com/
content/www/us/en/architecture-and-technology/ethernet/introducing-
800-series.html.
To see how the iWARP and RoCEv2 RDMA storage transport protocols
fared in our performance testing, visit http://facts.pt/gktwnar.
April 2021 | 6

Figure 6: Aerospike database read throughput performance, in millions of transactions per second, of the Dell EMC
PowerEdge R740xd at multiple server node counts with Intel Ethernet 800 Series Network Adapters with ADQ on and
ADQ off using Intel Optane persistent memory. Source: Principled Technologies.
Enabling ADQ with Intel Optane persistent memory
After testing with traditional DRAM memory alone, we also compared the performance effects of enabling ADQ
on a Dell EMC PowerEdge R740xd configured with Intel®
Optane™
persistent memory (PMem). Servers that use
2nd Generation Intel Xeon Scalable processors, like the Dell EMC PowerEdge R740xd, offer support for Intel
Optane PMem. As Figure 6 shows, enabling ADQ while using Intel Optane PMem showed performance gains
similar to those we saw in our testing with traditional memory, increasing read throughput by up to 52.9 percent.
Similar to DRAM testing, a single server node with Intel Optane PMem provided 29.1 percent lower weighted
average latency with ADQ enabled than with ADQ off. This shows that whether you use Intel Optane PMem or
DRAM, enabling ADQ can improve throughput, reduce latency, and increase response time predictability for
Aerospike databases on the Dell EMC PowerEdge R740xd compared to leaving ADQ disabled.
0
5,000,000
10,000,000
15,000,000
20,000,000
25,000,000
30,000,000
35,000,000
40,000,000
45,000,000
Aerospike database read throughput performance (with Intel Optane PMem)
ADQ-on ADQ-off
26,933,812
39,671,664
17,633,166
25,952,236
13,850,155
9,829,639
Transactions
per
second
(Higher
is
better)
Linear (ADQ-on) Linear (ADQ-off)
About Intel Optane persistent memory
Intel Optane persistent memory (PMem) is a new memory tier that offers both
performance and persistence, making it an excellent choice for hot, frequently read
data. Depending on your workload, you can use Intel Optane PMem in one of two modes:
• Memory Mode, where Intel Optane PMem acts as large-capacity, volatile memory,
with the system’s DRAM acting as a cache. In this mode, latency is slightly higher, but
capacity improves.
• App Direct Mode, where the applications and operating system are aware that two
types of load/store memory are configured on the system, and applications can choose
which type is best suited to meet the data’s latency and persistence requirements.
In our testing, we used App Direct Mode. To learn more about Intel Optane PMem,
visit https://www.intel.com/content/www/us/en/products/memory-storage/optane-dc-
persistent-memory.html
April 2021 | 7

Our tested server configurations with Intel Optane PMem provided comparable (a slight decrease) Aerospike
performance to using DRAM alone, while offering additional benefits, including data persistence and a larger
overall memory capacity per server node. Please note that in Aerospike testing, where you place the index and
the database affect performance results; for Intel Optane PMem tests, we placed the index on DRAM and the
database on PMem. For complete Intel Optane PMem testing results, including latency histograms, visit the
science behind the report.
Conclusion
Leveraging Intel Ethernet 800 Series Network Adapters with Application Device Queues in Dell EMC PowerEdge
R740xd servers can boost application throughput, lower latency, and increase predictability for your critical
Aerospike Database applications. As our testing with traditional memory shows, enabling ADQ can help you
make the most of your server resources by offering near-linear scaling and handling as many as 64.8 percent
more transactions per second. Similarly, configurations with Intel Optane PMem improved read throughput by
up to 52.9 percent when enabling ADQ. Turning on ADQ can help you shape your database traffic to deliver
improved performance for users, up to 32.0 percent lower average latency, and 41.7 percent better P99
predictability, providing greater consistency meeting SLAs.
1 Aerospike, “Aerospike Database 5,” accessed March 12, 2021, https://www.aerospike.com/products/database-platform/.
2 Intel, “Move Data Faster,” accessed September 24, 2020,
https://www.intel.com/content/www/us/en/architecture-and-technology/ethernet.html.
About the Dell EMC PowerEdge R740xd
What does the “xd” stand for? Extra drives—with up to 24 NVMe drives or 32 x 2.5” or 18 x 3.5” drives in just 2U of rack
space, the Dell EMC PowerEdge R740xd promotes flexibility through several drive options and density that allows you to use
software-defined storage solutions. The two-socket Dell EMC PowerEdge R740xd rack server features 2nd Generation Intel
Xeon Scalable processors, embedded Dell EMC iDRAC9, and Dell EMC OpenManage™
software for management. To learn
more about the Dell EMC PowerEdge R740xd, visit https://www.dell.com/en-us/work/shop/povw/poweredge-r740xd.
Principled Technologies is a registered trademark of Principled Technologies, Inc.
All other product names are the trademarks of their respective owners.
For additional information, review the science behind this report.
Principled
Technologies®
Facts matter.®
Principled
Technologies®
Facts matter.®
This project was commissioned by Dell EMC.
Read the science behind this report at http://facts.pt/8CTt1Ax
April 2021 | 8

Improve Aerospike Database performance and predictability by leveraging Intel® Ethernet 800 Series Network Adapters with Application Device Queues (ADQ)

Recommandé

Recommandé

Contenu connexe

Tendances

Tendances (20)

Similaire à Improve Aerospike Database performance and predictability by leveraging Intel® Ethernet 800 Series Network Adapters with Application Device Queues (ADQ)

Similaire à Improve Aerospike Database performance and predictability by leveraging Intel® Ethernet 800 Series Network Adapters with Application Device Queues (ADQ) (20)

Plus de Principled Technologies

Plus de Principled Technologies (20)

Dernier

Dernier (20)

Improve Aerospike Database performance and predictability by leveraging Intel® Ethernet 800 Series Network Adapters with Application Device Queues (ADQ)